Understanding consumer online shopping behaviour from the perspective of transaction costs

4.9.2 Model Specification 
Another issue with any form of SEM is the requirement for prior specification of the model 
(Kelloway 1995, Byrne 2009). The propositions incorporated in a model are most frequently 
drawn from previous research or theory (Bollen and Long 1993). In other words, pre-existing 
theoretical information is used to decide which variables (both observed and latent variables) 
to include in the proposed model, and how these variables are related. Model specification 
involves determining every relationship and parameter that is of interest to the researcher 
(Kelloway 1995, Byrne 2009). As highlighted by Schumacker and Lomax (2004), this is the 
hardest part of SEM as the exclusion of important variables or the inclusion of unimportant 
variables can produce an implied model that is misspecified resulting in bias parameter 
estimates (known as specification error). Specification error makes it likely that a theoretical 
model will not fit the data statistically and thus it will be deemed as unacceptable 
(Schumacker and Lomax 2004, Kline 2011). 

---

198
The structural model proposed and tested in this study draws on TCT (Williamson 1981b, 
Williamson and Ghani 2012) and prior research (Szajna 1996, Liang and Huang 1998, Teo
 et 
al.
2004, Teo and Yu 2005, Kim and Li 2009b). The measurement model is operationalized 
using multi-item scales specified according to factor structures used in previous research and 
the extant literature (Anderson and Gerbing 1988, Taylor and Todd 1995b, Pavlou and 
Chellappa 2001, Cho
 et al.
2003, Lee and Lin 2005, Teo and Yu 2005, Kim and Li 2009b, 
Wu
 et al.
2014)
. 
4.9.3 Model Identification 
In SEM, it is crucial that the level of model identification must be clarified prior to estimation 
of parameters (Schumacker and Lomax 2004). Traditionally, there have been three levels of 
model identification: underidentified, just-identified and overidentified (Bollen 1989, Byrne 
2009). According to Schumacker and Lomax (2004), a necessary but insufficient condition 
for model identification is ensuring that the number of free parameters to be estimated in a 
theoretical model must be less than (called overidentified) or equal to (called just-identified) 
the number of distinct values in the sample variance-covariance matrix (distinct values = 
p 
(
p 
+ 1)/2 where 
p 
is the number of observed variables). On the other hand if the number of free 
parameters exceeds the number of distinct values in the sample matrix (called 
underidentified), then the model parameters cannot be uniquely determined because there is 
not enough information in the matrix. A method for avoiding these problems is to ensure that 
either one indicator for each latent variable must have a factor loading fixed to 1, or that the 
variance of each latent variable must be fixed to 1 (Bollen 1989, Schumacker and Lomax 
2004, Byrne 2009). Utilizing this method will often eliminate the problem of scale 
indeterminacy; however, additional constraints (e.g. a parameter is constrained to equal one 
or more other parameters) may also be necessary (Schumacker and Lomax 2004). 

---

199